What Are Claude Connectors? Data Access, Auth, Custom Setup, and Troubleshooting (July 2026)

TL;DR: Claude Connectors let Claude use external apps and data through MCP. A connector can read data, take actions, and sometimes render interactive UI inside the conversation. Custom connectors use remote MCP servers that Claude reaches from Anthropic’s cloud. OAuth and source-system permissions decide whose data Claude can access. Test the MCP server, auth flow, tool contracts, and interactive UI before you connect it to a real Claude account.

What Are Claude Connectors?

Claude Connectors are integrations that give Claude access to external tools and data. Without a connector, Claude mostly sees the conversation, uploaded files, and host-provided context. With a connector, Claude can call a tool to search a document store, read a project issue, summarize recent messages, update a ticket, create a calendar event, or query an internal API.

For developers, a Claude Connector is best understood as an MCP boundary:

• The connector exposes tools with names, descriptions, schemas, and handlers.
• Claude decides when a tool is relevant to the user’s request.
• The tool returns model-readable content and structured data.
• Auth and permissions decide which user data the tool can reach.
• Interactive connectors can also return UI resources for Claude to render.

The connector is a product surface Claude can reason over and act through, so the tool descriptions, output shapes, permissions, and error states all matter.

Connector Types in 2026

Claude now has several connector paths, and the names can blur together.

Connector path	Where it runs	Best for
Directory web connector	Hosted service, enabled from the Connectors Directory	Reviewed SaaS and public integrations
Custom remote MCP connector	Your public HTTPS MCP server	Internal tools, proprietary apps, staging connectors
Desktop extension	Local Claude Desktop package	Local files, local tools, machine-specific workflows
API MCP connector	Remote MCP server used from the Claude API	Programmatic agent workflows
Interactive connector	Remote connector with MCP App UI resources	Dashboards, task boards, forms, previews, review screens

The practical split is remote versus local. Remote connectors are web-accessible MCP servers. Claude reaches them from Anthropic cloud infrastructure, even when the user opens Claude Desktop. Desktop extensions are a separate local mechanism for Claude Desktop and can reach local files or processes.

If you want a connector to work across Claude web, Claude Desktop, mobile, Cowork, and team workflows, start with a remote MCP server. If the connector is only useful on one user’s machine, a desktop extension may fit better.

What Data Claude Can Access

Claude can access the data your connector exposes through tools, limited by the connected user’s permissions. If the user cannot access a file, channel, record, project, or account in the source system, your connector should not return it.

There are four permission layers to think through:

1. Source-system permissions, such as the user’s access in Google Drive, Slack, Linear, Salesforce, or your own app.
2. OAuth scopes, which define what your connector may request from the source system.
3. MCP tool design, because each tool decides which read or write action exists.
4. Claude-side controls, where users or admins can enable, block, or require approval for tool categories.

This is why a connector should keep tools narrow. A search_docs tool with a query, result limit, and clear output shape is easier to review than a broad run_query tool that can read anything. A create_issue tool that names the project, title, body, and labels is easier to trust than a generic “write to app” tool.

Auth and Identity

Most production connectors need OAuth because Claude has to act with the user’s permission. OAuth gives your MCP server a user-scoped access token from the source system, so the connector can enforce the same access model that app already uses.

For custom remote MCP connectors, Claude lets you add OAuth Client ID and Client Secret in advanced settings. On Team and Enterprise plans, an owner or primary owner adds the connector to the organization, then individual users connect and authenticate it for their own accounts.

For development, a custom connector can be authless if it only exposes public or test data. Do not treat authless development as a production shortcut. If the server can read private customer data or perform write actions, add OAuth, narrow scopes, and log the important decisions.

For directory submission, expect reviewers to inspect the auth flow, requested scopes, privacy behavior, tool descriptions, write actions, and test account. A connector that asks for broad scopes or hides write actions behind vague tool names is harder to trust.

How to Add a Custom Connector

A custom Claude Connector is a remote MCP server Claude can reach over HTTPS.

Custom connectors using remote MCP are available on Claude, Cowork, and Claude Desktop for Free, Pro, Max, Team, and Enterprise plans. Free users are limited to one custom connector.

For an individual Pro or Max plan:

1. Deploy your MCP server or expose local development with a tunnel.
2. Open Claude settings and go to Customize > Connectors.
3. Click the plus button, then Add custom connector.
4. Enter the connector name and remote MCP server URL.
5. Add OAuth Client ID and Client Secret if the server requires auth.
6. Add the connector, connect it, then enable it in a conversation from the plus menu.

For Team and Enterprise plans, an organization owner adds the custom connector in Organization settings > Connectors. Members then find the configured connector, connect their own account, and enable it per conversation.

The network detail is easy to miss: Claude connects to your server from Anthropic’s cloud, not from your local machine. A server that works in your browser may still fail in Claude if it is behind a VPN, firewall, private DNS zone, or IP allowlist that excludes Anthropic’s ranges.

Web Connectors vs Desktop Extensions

Use a remote web connector when the integration should work across accounts, organizations, browsers, mobile clients, Cowork, or the API. It is the right default for SaaS products, internal company tools, and anything you may submit to the Connectors Directory.

Use a desktop extension when the integration needs local machine access. Examples include local files, local scripts, private developer tools, or workflows that should never leave the user’s computer. Desktop extensions are packaged and installed in Claude Desktop. They are useful, but they do not replace remote connectors for web and mobile use.

Claude Code is another path. It can use MCP servers directly rather than going through the same connectors UI. If you are building a connector for developers, test both the Claude connector path and the raw MCP server path because setup, auth, and logs differ.

Standard vs Interactive Connectors

A standard connector gives Claude tools. Claude calls the tool, receives data, and writes the answer in the conversation.

An interactive connector adds UI. The tool result can point Claude at an MCP App resource, and Claude can render that resource as an inline card or fullscreen view. This is useful when the user needs to inspect a table, edit a form, review a design, approve a change, compare records, or navigate more data than belongs in a text answer.

Build the standard connector first:

1. Define one clear read-only tool.
2. Return concise model-readable content and structured data.
3. Add OAuth and permissions.
4. Test the tool in edge states.
5. Add UI only where it improves the user’s task.

That order keeps the connector useful in headless MCP clients and in Claude flows where UI is not available. It also matches the broader MCP Apps direction. ChatGPT implements the open MCP Apps UI standard, and portable UI starts with the shared tool/resource contract before adding host-specific behavior.

Common Connector Problems

Custom connector cannot connect:

• Confirm the URL is public HTTPS and includes the correct MCP path.
• Test it from outside your company network.
• Check whether a firewall, VPN, WAF, or IP allowlist blocks Anthropic cloud traffic.
• Verify the MCP server supports remote transport and initializes cleanly.
• Use MCP Inspector to call tools and inspect schemas before retrying in Claude.

Auth works for one user but not another:

• Check source-system permissions first.
• Confirm OAuth redirect URIs and scopes match the deployed environment.
• Make sure Team or Enterprise owners enabled the connector for the organization.
• Check whether an admin blocked a write action in Claude.

Claude chooses the wrong tool:

• Rename vague tools.
• Add parameter descriptions that say what each field means.
• Split broad tools into focused read and write tools.
• Return clear error messages when a tool is used outside its scope.

Interactive UI does not render:

• Make sure the tool result points at the UI resource correctly.
• Check resource metadata, CSP domains, iframe-safe assets, and display modes.
• Keep the tool useful without UI so the workflow still works if the host cannot render the component.

Testing Before Shipping

Test a connector in layers.

Start with unit tests for tool handlers, schema validation, auth checks, source-system permission checks, and error mapping. Then run MCP-level tests that list tools, call tools, inspect resources, and validate structuredContent against outputSchema.

For interactive connectors, add rendered UI tests. Use deterministic tool results so you can test empty states, missing permissions, OAuth failures, cancelled work, long lists, write confirmations, and source-system errors without depending on a real host.

sunpeak packages that workflow for Claude Connectors and MCP Apps. The local inspector replicates Claude and ChatGPT runtimes, simulation files pin tool input and tool results, and Playwright tests can assert the rendered app state in CI. That lets you catch connector bugs before you spend time in a real Claude account.

Save live Claude tests for the things only Claude can prove:

• The deployed remote MCP server is reachable from Anthropic cloud.
• OAuth redirects and token refresh work in the real surface.
• Claude chooses the expected tools from real prompts.
• Interactive UI renders in the target display modes.
• Admin controls, write approvals, disconnect, and reconnect work.

Building Your Own Connector

You can build a Claude Connector with any MCP SDK and your normal server stack. The minimum production shape is:

• A remote MCP server reachable over HTTPS.
• Focused tools with clear names, descriptions, schemas, and output shapes.
• OAuth when the connector reaches private user data.
• Narrow write tools with confirmation-friendly descriptions.
• Logs that help debug initialization, auth, tool calls, and source-system errors.
• Tests for both the MCP server and the host-rendered UI, if the connector is interactive.

If you are building an interactive connector, sunpeak gives you the project structure, local Claude inspector, MCP App resource wiring, simulation fixtures, E2E tests, visual tests, live host checks, and CI workflow. The payoff is proof that the connector behaves across tool states, host runtimes, themes, display modes, auth states, and data shapes before users depend on it.

Start with the connector contract. Add UI when the data needs UI. Test the server and the rendered resource separately. Then run one narrow live Claude check before you submit or roll it out.